Repeated Median Velocity Strategy, Part 2 by Dennis Meyers, PhD
All Aboard The Trend!
Here in the second part of this article, you’ll identify optimal
parameters and find out how the system performed.
Last month in part 1, I described the repeated median
velocity (RMV) strategy and how it can be used to place
buy & sell orders as well as how to test this strategy.
Here in part 2, I will discuss how to find the optimized system
parameters so you can increase your overall trading profits.
Finding the system parameters
In the RMV strategy, there are three strategy parameters to
find: N, vup, and vdn. For the test data, I ran the TradeStation optimization engine on the Russell 2000 index emini futures
(TF) one-minute price bars from March 30, 2011 to May 3,
2013 with the following optimization ranges for the repeated
median strategy inputs:
* N from 20 to 70 in steps of 10
* vup from 0.5 to 10 steps of 0.5
* vdn from 0.5 to 10 in steps of 0.5.
I created 105 30-day in-sample periods each followed by a
seven-day out-of-sample period for the in-sample/out-of-sample
periods (see sidebar “Walk-Forward Out-Of-Sample Performance
Summary” at http://www.traders.com/files/MeyersSB.html). This will produce 2,400 different input combinations
or cases of the strategy input parameters for each of the 105
in-sample/out-of-sample files for the two years of one-minute
bar TF data.
What I am trying to do is statistically identify the best performance
metric (which I call a filter) or combination of best
performance metrics that I can apply to the in-sample section
that will give me strategy inputs that will produce, on average,
valid profits in the out-of-sample section, or future data.
When I run an optimization over many combinations of
inputs in TradeStation, it creates an output page with each
strategy input combination as its rows, and, as its columns,
various trading performance measures such as profit factor,
total net profits, and so on.
A simple filter would be to choose the strategy input optimization
row in the in-sample section that had the highest net
profit, or choose a row that had the best profit factor with its
associated strategy inputs. Unfortunately, this type of simple
metric performance filter has been found to rarely produce
good out-of-sample results. Some of the more complicated
metric filters have been found to produce good out-of-sample
results, as they minimize spurious price movement biases.